A backup nobody has restored is an unverified assumption.

Choose accelerators by framework fit; verify backups by actually restoring.

GPU

Broadest framework compatibility, general parallel compute

TPU

Google's own silicon, purpose-built for ML training/inference

A daily snapshot schedule "succeeded" for a year before a real restore revealed 5 months of backing up an empty placeholder disk after an undetected machine-type migration. Only a tested restore actually confirms a backup works.

A snapshot restores data; an image creates new bootable instances — different problems, even though you can convert between them. ZONE_RESOURCE_POOL_EXHAUSTED is often an availability problem, not a quota problem.